How to Build an AI Lead Qualifier for Service Businesses (Filter, Don't Just Follow Up)

This isn't the speed-to-lead pitch

It's worth drawing this line before anything else, because the two agents look similar from a distance and get pitched to the same buyer. A speed-to-lead agent solves for time: it replies in seconds so a lead doesn't go cold in an unanswered inbox. This guide solves for a different problem entirely — one that shows up in businesses that already reply fast, or that get more inquiries than they have capacity to serve: paid on-site estimators, contractors with a bounded service area, firms with intake requirements, anyone who has learned the hard way that not every inquiry is a fit.

For those businesses, the cost isn't a slow reply. It's hours spent driving to a job outside the service area, quoting a job type they don't do, or sitting on a call with someone who was never going to book. The qualifier's job is to let every lead feel answered — nobody gets ignored — while sorting the pipeline before it reaches a human, so the owner's time goes to the leads worth having. If you haven't read the speed-to-lead guide, start there for the instant-response case; this one picks up after the reply is already sent.

The spec

A qualifier is a short conversational intake, not an interrogation: what's the job, where is it, roughly when, and — only if the business actually prices this way — a budget band. Three to five questions, asked in whatever order the conversation naturally goes, not a rigid form marching through fields.

Every answer gets scored against the business's own criteria, not a generic template: service area (does the address fall inside the radius they'll actually drive), job type (do they do this kind of work at all), and minimums (is this job big enough to be worth a visit). That scoring produces a CRM tag — hot, nurture, or decline — and hot leads get an instant handoff to the owner exactly the way a speed-to-lead agent would hand one off, because qualification and speed aren't in competition; a good pipeline runs both.

The decline path is the part most builders skip and shouldn't: a lead outside the service area or job scope gets a graceful, specific answer — not a form silently going nowhere — ideally with a referral or an honest "that's not something we handle." And the guardrail that matters most here: the agent recommends a tag, it doesn't make the final call on anything ambiguous, and it never quotes a firm price. Scoring a lead hot or cold is a judgment call trained on the business's own criteria; deciding what to charge is the owner's call, always.

The build, honestly, both paths

DIY, this is a form or chat intake with a few questions, a set of scoring rules written in whatever logic your stack supports, and a write path into the CRM that tags the lead and files the answers. Connecting that intake to the tools it needs to read and write — the calendar, the CRM, the messaging channel — is exactly the integration problem the Model Context Protocol was built to reduce; the standard describes itself as letting an AI application "connect to data sources... tools... and workflows" without a custom integration per tool. None of the individual pieces are hard — the work is translating the owner's actual judgment ("we don't do jobs under $500," "we don't cross the river") into rules a script can apply consistently, and keeping those rules current as the business's criteria change.

That translation is the real asset here, more than the code around it. A written scoring-criteria doc — the client's judgment, put into words for the first time in a lot of cases — is worth more to the relationship than the automation wrapped around it, because it's the thing the owner can review, correct, and hand to a new hire later.

The assembled path is what we build: SeldonFrame wires intake, CRM tagging, and routing into one conversation, so the scoring criteria live alongside the agent rather than being reconstructed in a separate integration layer. SeldonFrame is our own product, so this particular recommendation comes from the interested party — discount it accordingly. Either path needs the same criteria doc; the difference is how much of the wiring is pre-connected.

Selling it: a different pain than missed calls

Don't pitch this with the speed-to-lead line — a business that already answers fast doesn't feel that pain. Ask a different question instead: "how much time did you spend last week on quotes that went nowhere?" Most owners who do paid estimates or field a lot of out-of-area calls have a number for that, and it's usually a number they've never said out loud to anyone selling them software.

The best-fit verticals follow from the problem: contractors who send someone out for a paid estimate, firms with real intake requirements (licensing, permits, insurance types), and any business with a hard service-area boundary. If the owner has ever said "I wish people would stop calling about jobs we don't do," that's the opening line for this pitch, not the speed-to-lead one.

Pricing and running it

Same shape as most agent retainers: a monthly fee, with a report the owner actually reads. For a qualifier, that report is leads processed, hot-lead response time (so the qualification layer doesn't quietly become the new bottleneck), and an honest, clearly-labeled estimate of owner-hours saved — never a hard number presented as measured fact, since there's no way to observe the counterfactual hours a business would have spent without the agent.

The criteria drift, and the retainer should account for that: sit down with the client quarterly and walk through what got tagged hot, nurture, and decline over the period, and adjust the rules to match what they'd have actually wanted. A qualifier tuned once at setup and never revisited slowly drifts out of sync with how the business actually operates.

Failure modes to design against

Over-qualifying is the most common one: an agent that keeps asking questions until a genuinely hot lead gets annoyed and leaves. Cap the question count and bias toward handing off early when a lead is clearly interested — qualification should never cost you a lead speed-to-lead would have caught.

Scoring on criteria the client never actually agreed to is the second: a rule you inferred rather than one the owner confirmed in the criteria doc will eventually tag the wrong lead the wrong way, and the client will find out from a customer complaint instead of from you.

Third, silently declining leads the owner would have wanted is worse than it sounds, because nobody notices until a competitor books the job. Log every decline with the reason, and review the log with the client — don't let the decline path run as a black box.

Fourth, and the general rule underneath the other three: the false-positive trap. It's better to pass a maybe to the human than to auto-decline it. A qualifier that's too eager to say no costs the business real revenue in a way that's much harder to notice than a qualifier that's a little too generous with hot tags.

Frequently asked questions

Sources

• Model Context Protocol — "What is the Model Context Protocol (MCP)?"
• Twilio — SMS Pricing (US)